cellRank builds a graph based on a transition matrix. The transition matrix contains the probabilities of cell–>cell transitions, which are calculated as a weighted average between (similarity of velocity vector and cell difference vector) and (euclidean similarity in PCA space). Here, we want to compare this graph to the graph that we build using projectedNeighbors.

Setup and get data from scVelo

Use the reticulate package to use scVelo/cellrank from within R:

scVelo - whole trajectory

Build graph with all data i.e. with cells from all parts of the trajectory.

load complete adata from previous run

load pancreas data

Run scVelo

Run velocity

Run cellrank

Save adata to speed up later

Extract transition matrices

## <Conn[dnorm=True]>
## <Velo[softmax_scale=3.8, mode=deterministic, seed=3185]>

Weighted graphs from transition matrices
adjacency = transition matrix –> fdg
Distances: transition matrix using just euclidean similarity
Velocities: transition matrix using just velocity correlation
Combined: transition matrix = weighted average of distance and velocity
(using default weight 0.8velocity + 0.2distance)

Varying weight in weighted average

scVelo - missing intermediates

Build graph with data where part of the trajectory is missing.

load missing intermediate adata from previous run

load pancreas data

Run scVelo

Run velocity

Run cellrank

Save adata to speed up later

Extract transition matrices

## <Conn[dnorm=True]>
## <Velo[softmax_scale=3.86, mode=deterministic, seed=65033]>

Weighted graphs from transition matrices
adjacency = transition matrix –> fdg
Distances: transition matrix using just euclidean similarity
Velocities: transition matrix using just velocity correlation
Combined: transition matrix = weighted average of distance and velocity
(using default weight 0.8velocity + 0.2distance)

Varying weight in weighted average

veloViz - whole trajectory

Extract count data..

Filter genes

(Downsample cells to make things easier)

Normalize for dimensional reduction

## Warning in if (!class(counts) %in% c("dgCMatrix", "dgTMatrix")) {: the condition
## has length > 1 and only the first element will be used
## Converting to sparse matrix ...

Dimensional reduction

Run velocyto on panc data

## Warning in if (!class(counts) %in% c("dgCMatrix", "dgTMatrix")) {: the condition
## has length > 1 and only the first element will be used
## Converting to sparse matrix ...

Scores of observed and projected states in PC space

Graph visualization on subset of cells from PC coordinates

veloViz - missing intermediates

Extract count data..

Filter genes

(Downsample cells to make things easier)

Normalize for dimensional reduction

## Warning in if (!class(counts) %in% c("dgCMatrix", "dgTMatrix")) {: the condition
## has length > 1 and only the first element will be used
## Converting to sparse matrix ...

Dimensional reduction

Run velocyto on panc data

## Warning in if (!class(counts) %in% c("dgCMatrix", "dgTMatrix")) {: the condition
## has length > 1 and only the first element will be used
## Converting to sparse matrix ...

Scores of observed and projected states in PC space

Graph visualization on subset of cells from PC coordinates

Changing veloviz parameters

Above, I used k=30 for direct comparison to cellRank graph which computes distances to K=30 nearest neighbors uses. However, this might not be where veloviz performs best.

Comparing graphs

Compare mean distance between cells before and after the gap normalized by max distance between any two cells for each graph.
Comparing default cellRank graph to veloViz graph

Comparing ‘best’ cellRank graph (only velocity based) to ‘best’ veloViz graph (k = 10)

Comparing graphs from each method